Scholarlink

<p>Neonatal seizure detection is crucial for preventing neurological damage, as these seizures are linked to abnormal brain activity. While machine learning and deep learning show promise for EEG analysis, persistent challenges remain, including signal non-stationarity, class imbalance, and limited interpretability. We propose a hybrid deep framework for early neonatal seizure detection. The method uses discrete wavelet transform (DWT) to decompose EEG signals into subbands and short-time Fourier transform (STFT) to create time-frequency spectrograms. Inception-ResNetV2 extracts deep features from these spectrograms, which are then classified by an XGBoost model. This model is fine-tuned offline using particle swarm optimization (PSO) to enhance performance. This hybrid approach tackles key challenges by combining powerful deep representations with a fast, explainable classifier. This design specifically addresses signal non-stationarity and class imbalance while minimizing computational load and preventing overfitting. The framework’s performance was evaluated on the Helsinki Neonatal EEG Dataset, which contains data from 79 term neonates, including 460 expert-annotated seizures. Our model achieved an average accuracy of 98.75%, with 98.56% precision, 98.36% sensitivity, and 98.91% specificity, outperforming most existing methods. Using medium analysis windows further improved accuracy to 99.82%, with some cross-validation runs reaching 100%. Subband analysis highlighted the importance of theta and delta bands in the frontotemporal regions for early detection. The integration of Inception-ResNetV2 and PSO-optimized XGBoost provides significant performance gains while offering insights into the role of EEG subbands in seizure manifestation. This framework is a practical solution for neonatal intensive care unit (NICU) monitoring, offering a robust foundation for a clinical decision-support system to enable rapid and accurate seizure detection.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Neonatal seizure detection from EEG using inception ResNetV2 feature extraction and XGBoost optimized with particle swarm optimization

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年11月24日

查看全文
  • Nazanin Nemati,
  • Saeed Meshgini,
  • Tohid Yousefi Rezaii,
  • Reza Afrouzian

摘要

Neonatal seizure detection is crucial for preventing neurological damage, as these seizures are linked to abnormal brain activity. While machine learning and deep learning show promise for EEG analysis, persistent challenges remain, including signal non-stationarity, class imbalance, and limited interpretability. We propose a hybrid deep framework for early neonatal seizure detection. The method uses discrete wavelet transform (DWT) to decompose EEG signals into subbands and short-time Fourier transform (STFT) to create time-frequency spectrograms. Inception-ResNetV2 extracts deep features from these spectrograms, which are then classified by an XGBoost model. This model is fine-tuned offline using particle swarm optimization (PSO) to enhance performance. This hybrid approach tackles key challenges by combining powerful deep representations with a fast, explainable classifier. This design specifically addresses signal non-stationarity and class imbalance while minimizing computational load and preventing overfitting. The framework’s performance was evaluated on the Helsinki Neonatal EEG Dataset, which contains data from 79 term neonates, including 460 expert-annotated seizures. Our model achieved an average accuracy of 98.75%, with 98.56% precision, 98.36% sensitivity, and 98.91% specificity, outperforming most existing methods. Using medium analysis windows further improved accuracy to 99.82%, with some cross-validation runs reaching 100%. Subband analysis highlighted the importance of theta and delta bands in the frontotemporal regions for early detection. The integration of Inception-ResNetV2 and PSO-optimized XGBoost provides significant performance gains while offering insights into the role of EEG subbands in seizure manifestation. This framework is a practical solution for neonatal intensive care unit (NICU) monitoring, offering a robust foundation for a clinical decision-support system to enable rapid and accurate seizure detection.